Yzaza/nih-cxr-models

NIH ChestX-ray14 multi-label classifiers

Multi-label classifiers for the 14 ChestX-ray14 pathologies, trained on the official patient-wise split (75,312 train_val / 25,596 test) so results are comparable with the published benchmark table.

Models

file	architecture	resolution	val AUROC
densenet121_wbce_asl_medical_320.pt	densenet121	320px	0.8414
swin_t_wbce_asl_medical_320.pt	swin_t	320px	0.8395

Test results (official split, patient-level bootstrap)

model	mean AUROC	95% CI	mean AUPRC
densenet121_wbce_asl_medical_320	0.8106	0.8036-0.8167	0.2855
swin_t_wbce_asl_medical_320	0.8184	0.8127-0.8236	0.2818

Training setup

• Loss: weighted BCE combined with Asymmetric Loss (gamma_neg=4, gamma_pos=0, clip=0.05); the label matrix is ~94.8% negative.
• Augmentation: rotation +/-12 deg, translation +/-8%, scale 0.92-1.08, shear +/-5 deg, brightness/contrast +/-12%. Horizontal flip is deliberately excluded: mirroring a chest radiograph produces anatomically invalid images (dextrocardia) and measurably degraded laterality-dependent findings.
• Optimiser: AdamW, cosine schedule with warmup, discriminative LR (head at 10x backbone).
• Selection: early stopping on validation macro AUROC.

Labels

[
  "Atelectasis",
  "Cardiomegaly",
  "Consolidation",
  "Edema",
  "Effusion",
  "Emphysema",
  "Fibrosis",
  "Hernia",
  "Infiltration",
  "Mass",
  "Nodule",
  "Pleural_Thickening",
  "Pneumonia",
  "Pneumothorax"
]

Usage

import torch, torchvision.models as tvm, torch.nn as nn
from huggingface_hub import hf_hub_download

p = hf_hub_download("Yzaza/nih-cxr-models", "swin_t_320.pt")
ck = torch.load(p, map_location="cpu", weights_only=False)
m = tvm.swin_t(weights=None)
m.head = nn.Linear(m.head.in_features, 14)
m.load_state_dict(ck["model"]); m.eval()
# preprocess: grayscale -> 3ch, resize to ck["crop"], ImageNet normalisation

Intended use

Research demonstration only. Trained on NLP-mined labels with a documented error rate; not a medical device and not validated for clinical use.